A Systematic Review of Reimagining Fashion and Textiles Sustainability with AI: A Circular Economy Approach

Featured Application

This work seeks to highlight the current AI applications within the Fashion and Textiles industry to support a transition to a Circular Economy.

Abstract

Artificial intelligence (AI) is revolutionizing the fashion, textile, and clothing industries by enabling automated assessment of garment quality, condition, and recyclability, addressing key challenges in sustainability. This systematic review explores the applications of AI in evaluating clothing quality and condition within the framework of a circular economy, with a focus on supporting second-hand clothing resale, charitable donations by NGOs, and sustainable recycling practices. A total of 135 research resources were identified through searching academic databases including Google Scholar, Springer, ScienceDirect, IEEE, Taylor and Francis, and Sage journals. These publications were subsequently refined down to 49 based on selected inclusion criteria. The selection of these sources from diverse databases was undertaken to mitigate any potential bias in the selection process. By analyzing the effectiveness and challenges of related peer-reviewed articles, conference papers, and technical reports, this study highlights state-of-the-art methodologies such as convolutional neural networks (CNNs), hybrid models, and other machine vision systems. A critical aspect of this review is the examination and analysis of datasets used for model development, categorized and detailed in a comprehensive table to guide future research. Whilst the findings emphasize the potential of AI to enhance quality assurance in second-hand clothing markets, streamline textile sorting for donations and recycling, and reduce waste in the fashion industry, they also highlight gaps in the available datasets, often due to limited size and scope. The types of textiles captured were most commonly swatches of fabric, with 20 studies examining these, whereas whole garments were less frequently studied, with only 7 instances. This review concludes with insights into future research directions and the promising use of AI within fashion and textiles to facilitate a transition to a circular economy. This project was supported through RMIT University’s School of Fashion and Textiles internal seed funding (2024).

1. Introduction

The use of artificial intelligence (AI) technology within the fashion and textiles industry is growing, with potential applications often focusing on consumer experience (e.g., virtual try-on and digital avatars) to consumer insights, forecasting, and trend analysis [1]. AI is a powerful tool that brands may employ to drive sales, improve consumer experience, and increase efficiency [2]. However, there is also enormous scope for the use of AI within the industry to drive and develop new sustainability initiatives, divert waste, and re-circulate secondhand goods.

The global fashion and textile (F&T) industry is a significant contributor to environmental pollution and degradation, largely driven by the fast fashion model that promotes rapid production and consumption cycles. The exact volume of textile waste this sector generates annually is unknown; 92 million tons is frequently referenced, much of which is either incinerated, exported or ends up in landfills, exacerbating the environmental crisis [3,4]. Currently, these pathways are challenged as export markets are becoming oversaturated [4], with the volumes of textiles being exported, dumped, and landfilled beginning to exceed capacities [5]. Economically, the premature disposal of clothing represents a loss exceeding USD 400 billion annually, highlighting inefficient resource use and missed recycling opportunities [2]. Socially, developing countries often receive large quantities of second-hand clothing, overwhelming local markets and contributing to waste management challenges and public health issues [4]. Environmentally, the decomposition of synthetic fabrics in landfills releases harmful chemicals, exacerbating pollution and environmental injustice [4]. Addressing these challenges requires innovative approaches that extend the lifecycle of textiles and garments, emphasizing reuse, repair, and recycling infrastructures and technology. A circular economy (CE) framework offers a promising solution, aiming to minimize waste and resource usage by creating closed-loop systems [3].

Given the urgency of the F&T waste crisis, previous studies have undertaken systematic reviews on textile waste and circular economy frameworks. However, the focus has primarily been on a more holistic and critical view of the fashion production system, its complex and global supply chain, and assessing existing technologies for material and energy recovery [5] or recycling [6]. While sorting textile waste is only one of the barriers to transitioning to a CE, it is a large one, as identified by Chen et al. [6] and many others.

The circular economy concept aims to redefine growth by decoupling economic activity from the consumption of finite resources and designing waste out of the system. In the context of textile recycling and reuse, CE goals emphasize reducing textile waste by extending the lifecycle of products through recycling, reusing, and repurposing materials. This approach not only conserves resources but also reduces environmental impact and fosters sustainable development. By integrating AI-driven textile sorting solutions, we can enhance the efficiency and accuracy of material recovery processes, making significant strides towards achieving CE goals and promoting a more sustainable textile industry.

Within this context, AI has emerged as a powerful tool for assessing garment quality and condition, transforming traditional practices and supporting sustainable initiatives [7]. AI technologies, particularly machine learning (ML), deep learning (DL), and computer vision (CV), have demonstrated significant potential in addressing key challenges in the fashion and textile industry. These technologies enable the automation of garment quality inspections, defect detection, textile sorting, and degradation prediction, offering speed, accuracy, and scalability [8]. For instance, recent studies [9,10] have shown that AI can significantly enhance the reliability of garment inspections while reducing labor cost. One of the most pressing needs in the circular economy is the effective sorting of textiles for recycling and reuse. Currently, much of this sorting is accomplished manually, especially in countries like Australia, where the charitable reuse sector relies heavily on volunteer labor. Existing technological systems for garment and textile sorting, such as near-infrared spectroscopy, have primarily been explored for use within commercial textile sorting organizations to support fiber separation for textile recycling [10]. As a result, it is unclear the extent to which these technologies are viable for the charitable or not-for-profit reuse sector, which is characterized by a subjective, often volunteer workforce who typically conduct localized value-based assessments around brand value, garment archetype, or current trend alignment. AI-driven systems, such as those employing convolutional neural networks (CNNs) and hybrid models, can classify textiles by type, physical condition, and recyclability, addressing a critical bottleneck in textile waste management.

Beyond recycling, AI has the potential to play a transformative role in second-hand clothing markets. Automated systems can be used to assess garment quality, detect defects, and estimate wearability, enabling accurate pricing and categorization. This not only enhances consumer trust but also supports charitable organizations in distributing usable garments effectively in the future.

This systematic literature review (SLR) aims to provide a comprehensive understanding of the current landscape of existing AI technology that is relevant to sorting textiles and garments and explores how these technologies can be implemented to improve sorting practices. While previous systematic literature reviews have examined the potential for use of AI technologies in the fashion industry [1,2], these reviews included a wide variety of AI technologies and applications, including the design and sale of goods. Study [1] reviewed the role of artificial intelligence (AI) in enhancing sustainability within the fashion industry from 2010 to 2022. It highlights AI’s diverse applications, including supply chain management, creative design, and waste control, emphasizing its potential to improve environmental sustainability. This study acknowledges the benefits of AI, such as efficient resource management and customer satisfaction, while also noting challenges like data requirements and implementation costs. Overall, AI is seen as a crucial tool for guiding the fashion industry towards a more sustainable future, despite some existing limitations. Study [2] systematically reviewed circular economy strategies in the textile industry. It identifies key strategies such as reuse, recycling, repair, and reduction, evaluating them through environmental, social, and economic lenses. The study highlights the potential of reuse to reduce waste and social inequality, while recycling faces technological and policy barriers. Repair extends garment lifespans but is limited by service accessibility and consumer knowledge. Reduction focuses on sustainable materials, challenged by the fast fashion model. Overall, the review underscores the need for substantial investments and supportive policies to advance sustainability in the textile industry.

However, a systematic review examining existing AI and machine vision technologies that could support the fashion industry to transition to a circular economy, with a focus on existing AI technologies for the purpose of sorting textile waste and discarded garments, does not appear to have been undertaken. By examining the recent academic literature, this unique review aims to remedy this gap by highlighting the potential of AI to support the secondhand textile industry, promoting CE strategies through increasing reuse of existing textiles and garments.

The principal research question this SLR seeks to answer is “What are the potential applications of computer vision for supporting the sorting and classification of textiles and garments to promote a more circular fashion industry?” This SLR provides the materials and methods utilizing the PRISMA methodology and provides a detailed discussion on the included papers and sources to discuss the AI application on textiles and their impact including assessment of dataset availability and practicality. In addition, it discusses the societal and economic impacts and articulates challenges and limitations on CE transition applications. This SLR then provides future directions and concludes with providing an answer to the posed research question.

2. Materials and Methods

This systematic review followed the guidelines outlined in the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) framework [11,12], which is widely recognized as a standard for reporting systematic reviews. The subsequent sections detail the steps undertaken throughout the review process.

2.1. Aim

This SLR aims to explore how computer vision technologies can be leveraged to enhance sustainability in the fashion industry by improving sorting processes related to recycling, reusing, and repurposing textiles and garments that are considered ‘waste’.

2.2. Eligibility Criteria

2.2.1. Inclusion Criteria

The following studies were included in this review:

• Research focus: identification and analysis of physical or visual indicators of garment aging, such as color fading, pilling, surface abrasion, and seam damage;
• Garments defect detection: detection of quality defects in finished textile products, including common issues like stains and stitching weaknesses using AI-driven methods or computer vision;
• Textile sorting: examination of AI techniques applied to the sorting of textiles in both pre-consumer (manufacturing waste) and post-consumer (used clothing) stages;
• Sustainability and recycling: automated textile sorting systems aimed at improving sustainability and recycling practices through AI;
• Dataset introduction: research articles introducing new datasets for second-hand clothing or damaged garments;
• Publication quality: inclusion of only peer-reviewed journal articles, conference papers, and technical reports from reputable sources (research centers) in textile technology, AI applications, or sustainability-focused research;
• Studies published between 2016 and 2024 were included to ensure the review reflects the most current trends, methods, and findings relevant to the topic;
• Language: only studies written in English were included in this review;
• Full text accessibility: only studies with the full text accessible were included, allowing for a thorough evaluation of the methodology, results, and relevance of each study.

2.2.2. Exclusion Criteria

The following studies were excluded from this review:

• Non-textile sorting: AI-based waste sorting techniques not specifically targeted at textile or garment sorting;
• Hardware development: studies focused on hardware development or engineering specific to textile waste sorting;
• Binary classification models: papers utilizing only binary classification models (e.g., defect/no defect) without specifying or categorizing the types of garment defects;
• Production stage defects: research applying AI for detecting defects at the textile production stage unrelated to finished, wearable garments;
• Production-focused metrics: studies using AI to predict production-focused quality metrics without direct implications for garment aging, wearability, or consumer use;
• Language: articles written in languages other than English were excluded;
• Review papers: review articles were excluded from the analysis to focus solely on original research;
• Studies published before 2016 or after 2024 were excluded to maintain a focused and contemporary literature scope.

2.3. Information Sources

A range of scholarly sources were identified to articulate the relevant research initiatives associated with the investigation of this question (

Table 1. Keywords and details of applying exclusion and inclusion criteria.

Keywords	“Post-consumer textile sorting” OR “Pre-consumer textile sorting” OR “Textile sorting” OR “fabric defects detection” OR “Garments defect detection” OR “Damage Detection in Circular Fashion” AND “Artificial Intelligence” OR “Computer vision” OR “Automatic” OR “Deep learning” OR “AI” OR “Machine learning” OR “Expert systems”
Timespan/Filter	2016–2024
Search systems	Google Scholar, Springer, ScienceDirect, IEEE, Taylor and Francis, Sage Journals
Criteria	Sources	No. of exclusion	No. of inclusion
Article type	Journal articles	61	37
	Conference papers	25	10
	Master’s/doctorate thesis	0	2
Language	English	0	49

). These sources were accessed through academic databases including Google Scholar, ScienceDirect, IEEE, Taylor and Francis, Springer, and Science Direct. The selection of these sources from diverse databases was undertaken to mitigate any potential bias in the selection process. The compiled sources exclusively encompass published materials to ensure the integrity and quality of this review. These sources were predominantly restricted to scholarly research that presents original contributions to the field. The classification process refined the scholarly research to include only peer-reviewed and cited academic papers (see Figure 1).

2.4. Search Strategy

The terms and keywords analyzed in this review encompass the following: post-consumer textile sorting, pre-consumer textile sorting, textile sorting, fabric defect detection, garment defect detection, damage detection in circular fashion, artificial intelligence (AI), computer vision, automation, deep learning, machine learning, and expert systems. Boolean syntax was also utilized to these terms to create optimal combinations using AND and OR operators to identify the most pertinent literature. The keywords were searched within the titles of the academic resources. We used the same set of search queries across all databases to ensure consistency in the article selection process. The databases were searched in November 2024. In addition to our primary database searches, we utilized Google Scholar to identify relevant studies not indexed in traditional academic databases. This platform enabled access to a wide array of scholarly materials, including grey literature and theses, thereby broadening the scope of our review. Furthermore, we conducted citation tracking by examining the reference lists of key articles (backward citation tracking) and identifying newer studies that cited these works (forward citation tracking). This approach facilitated the discovery of influential research and provided insights into the development of ideas within the field.

2.5. Selection Process

All initially selected articles underwent manual review to eliminate duplicates. Articles that were inaccessible—specifically, those without available full-text—were excluded. Following this, titles and abstracts were reviewed, and full-text screening was conducted, during which inclusion and exclusion criteria were applied to determine eligibility. Ultimately, articles that satisfied the inclusion criteria and aligned with this study’s objectives were included in the final review.

2.6. Data Collection Process

A data extraction sheet was developed for this stage of this study. An initial version was created to begin the process. To facilitate data collection, the online automatic information extraction tool SciSpace (https://scispace.com/ (accessed on 12 December 2024)) was used to extract relevant details from each study, including the problem statement, techniques used, datasets, and results. All extracted information was reviewed and verified by the researchers, with additional modifications made to the sheet, including the addition of necessary columns. Finally, the revised extraction sheet was validated to ensure it functioned correctly and effectively captured all relevant information. All retrieved results were independently verified by all researchers.

2.7. Data Items

To address our research questions, the selected articles were carefully analyzed to extract the following key information:

• Aim of the work;
• Technique used;
• Country of the article;
• All relevant information about dataset:
• Accessibility: public or not;
• Source: created or existing;
• Size: number of images;
• Image information: imaging method and image size;
• Textile type: size of swatch or garment captured in images;
• Publication year;
• Article type.

Relevant information within each dataset was selected in response to the research question determined by the interdisciplinary team. AI methods and dataset information were determined by HN and AA, while the F&T research team, including JE, RVA, GM, and SI, determined textile information categories. Information about the imaging method and size was inconsistently available and was not included in the final report. When information was unavailable, this was noted as ‘unclear’ or ‘not applicable’ as appropriate.

2.8. Synthesis Methods

The data collected were organized in Microsoft Excel and cleaned, with categories clarified and determined in discussion with all authors. This included refining categories for areas such as dataset origins and types of textiles to quantitatively assess them. Based on refined categorizations, the excel file was cleaned, with tables and graphs generated as a result. A tree diagram, line charts, and column graphs were used to represent and examine the data, supporting discussion between all researchers to identify emerging themes and topics. This study used quantitative methods to compare the types of datasets used and drew on qualitative methods to thematically analyze the studies and identify the potential applications of AI for supporting the sorting and classification of textiles and garments to promote a more circular fashion industry. Due to the nature of the dataset, and the diversity across the studies regarding the types of textiles and use of AI, it was determined that statistical analysis was not applicable.

2.9. Risk of Bias

To ensure the quality and reliability of the evidence presented in this review, we applied the ROBIS (Risk of Bias in Systematic Reviews) tool to assess potential sources of bias throughout the review process. The assessment focused on four key domains: study eligibility criteria, identification and selection of studies, data collection and appraisal, and synthesis and findings. For each domain, we addressed the relevant signaling questions using the standardized ROBIS response options (Yes, Probably Yes, Probably No, or No), and we derived the level of concern based on the cumulative responses, in accordance with the ROBIS guidelines. The outcomes of the ROBIS risk of bias assessment are summarized in

Table 2. ROBIS risk of bias assessment summary: this table presents the domain-wise evaluation of potential bias in this review using the ROBIS tool. Each domain was assessed based on signaling questions, with concern levels determined in accordance with ROBIS guidelines.

Domain	Risk of Bias Judgement	Rationale for Concern
1. Concerns regarding specification of study eligibility criteria	Low	This review clearly outlined pre-specified inclusion and exclusion criteria that were well-aligned with the review objectives.
2. Concerns regarding methods used to identify and/or select studies	Low	The search strategy was systematic and involved multiple reputable databases, ensuring broad coverage of relevant literature. The process for both screening titles and abstracts and assessing full-text papers included multiple reviewers, reducing the likelihood of selection bias.
3. Concerns regarding collecting data and appraise studies	Low	Data extraction was conducted systematically using a structured form that ensured consistency across studies. Key methodological and technical characteristics were captured in detail. All articles were assessed independently by a minimum of two reviewers, and appropriate data were abstracted independently, minimizing the potential for bias and error.
4. Concerns regarding the synthesis and findings	Low	The findings were systematically synthesized according to the application domains—textile sorting, garment defect detection, and piling identification. This structure allowed for the identification of key trends, commonly used methods, and differences across each area. Limitations of individual studies, such as limited dataset availability, lack of image diversity, restricted real-world validation, and model complexity, were explicitly acknowledged and discussed.

, which provides a domain-wise evaluation of potential bias and supports the overall credibility of the review findings.

3. Results

A total of 187 articles were initially identified from the selected online sources. Following the screening and selection process outlined in Figure 1, 49 studies met the inclusion criteria and were ultimately included in this review.

Figure 2a categorizes the sources of research, showing that most studies were published in peer-reviewed journals, reflecting the high academic rigor and credibility of this area of investigation. Figure 2b illustrates the trend in publications over time, showing a steady rise as the years progress. Research activity reaches its peak in 2023 with 14 publications, likely driven by the growing availability and capabilities of AI technologies. This pattern reflects a significant surge in interest after 2020, fueled by advancements in AI and a heightened focus on sustainability and efficiency in the textile industry. The stabilization observed in recent years suggests that foundational research in this field may be maturing, potentially signaling a shift toward practical implementation and real-world applications for sorting and defect detection.

Figure 3 shows that IEEE Xplore, Elsevier, and MDPI were the leading sources of articles in this review, each contributing a significant share and together accounting for over half of the total publications. Taylor & Francis (15.2%) and Sage Journals (10.9%) also made notable contributions, indicating their relevance to the review’s focus. The “Other” category, which includes theses, also accounted for 10.9%, reflecting the inclusion of grey literature in the analysis. All remaining sources collectively made up the rest of the contributions, representing a smaller share of the overall distribution.

Figure 4 provides a breakdown of research papers based on the countries associated with the first author’s institution. China leads significantly with 25 papers, followed by South Korea, Portugal, and Spain, each contributing 3 papers. Other countries, such as the United States, Taiwan, and India, had smaller contributions, while several nations, including Pakistan, Finland, and Sri Lanka, have a single paper each. This distribution highlights the dominant role of Chinese researchers in this field and the diverse, albeit smaller, contributions from other countries.

To present a structured synthesis of the reviewed literature, we divided the studies into three distinct application areas and organized them accordingly into

Table 3. Overview of studies utilizing AI techniques for textile sorting, highlighting the year, study focus, applied techniques, number of images, dataset type, and image source.

Paper	Year	Study Focus	Technique	No: Images	Dataset Type	Dataset Source	Fabric/Fiber Type
[10]	2021	Textile material recognition	NIR	253	fabric (section)	created	Cotton, polyester, viscose, and blended
[13]	2024	Garments type classification	CNN	26,833	garment (full)	created	11 garment categories, e.g., shirt, jacket, etc.
[14]	2021	Recycled clothing classification	Alex Net	3300	garment (full)	existing	9 garments categories
[15]	2023	Automated color sorting of multi-colored waste textiles	Object detection and decision tree	277	garment (full)	created	Not given
[16]	2016	Textile material classification	SIMCA, SVM, and ELM	120	fabric (section)	created	Natural and synthetic
[17]	2020	Natural vs. synthetic fiber identification	PCA, CVA, and k-NN	350	fabric (section)	created	Cotton, linen, wool, silk, viscose, polyamide, and synthetic fiber
[18]	2021	Textile material recognition	Neural network	892	fabric (swatch)	created	Polyester, cotton, wool, viscose, nylon, silk, acrylic, and blended
[19]	2022	Identification of textile fibers	CNN vs. (KNN and SVM)	600,404	yarn (detail)	created	25 textile fiber types
[20]	2023	Classification of fabric material	Supervised classification	104	multiple	created	Cotton, polyester, wool, silk, and viscose.
[21]	2024	Fabric identification	Deep learning	840	multiple	created	14 types of fabric
[22]	2023	Multi-classification of textile waste	Hybrid CNN-LSTM	10,000	fabric (section)	mixed	Not given
[23]	2020	Classification of textile waste	CNNs	263	fabric (swatch)	created	Polyester, wool, cotton, nylon, and blended
[24]	2022	Classification of fabric material	CNNs	370	fabric (unspecified)	created	Cotton, linen, wool, silk, polyester, polyamide, and viscose
[25]	2023	Identification of textile materials	AONet	7000	fabric (microscopic)	existing	Cotton, denim, linen, nylon, and silk
[26]	2022	Textiles material Recognition	CNN	2764	fabric (draped)	created	Polyester, cotton, wool, silk, viscose, nylon, acrylic, and blended
[27]	2023	Cotton percentage prediction in fabric	CNN	7000	fabric (swatch)	existing	Fabric with cotton percentage in it
[28]	2022	Color classification of waste textiles	CV	2,100,466	fabric (draped)	created	Not given

,

Table 4. Overview of studies utilizing AI techniques for defect detection, highlighting the year, study focus, applied techniques, number of images, dataset availability, dataset type, and defect types. The dataset specifies whether the images used for training AI models representing full garments or fabric sections of varying sizes.

Paper	Year	Study Focus	Technique	No: Images	Dataset Availability	Dataset Type	Dataset Source
[29]	2023	Fabric defect detection	YOLOv5	647	public	garment (full)	created
[30]	2023	Garment defect detection	CNN	800	private	garment (detail)	created
[31]	2024	Fabric defect detection	YOLOv8	2800	public	fabric (swatch)	existing
[32]	2024	Fabric defect detection	GSL-YOLOv8n	6345	public	fabric (swatch)	existing
[33]	2022	Garment inspection	Image processing	174	private	garment (full)	created
[34]	2024	Zipper tapes defect detection	CNN and YOLOv5	1200	private	trims (detail)	existing
[35]	2023	Clothing inspection for visually impaired	CNN	11,604	public	garment (full)	mixed
[36]	2023	Garments defect detection	YOLO-SCD	7021	public	fabric (swatch)	existing
[37]	2024	Fabric sewing break detection	U-Net Network	1000	private	garment (detail)	created
[38]	2022	Broken stitch detection	CNN	28	private	garment (detail)	created
[39]	2021	Zipper tape defect detection	CNN	11,201	private	trims (detail)	created
[40]	2024	Garments damage detection	CV	16,066	public	garment (full)	created
[41]	2023	Garments defect detection	YOLOv5	3161	private	garment (detail)	created
[42]	2022	Sewing stitch detection	DeepLabV3+ EfficientNet	900	private	garment (detail)	created
[43]	2020	Fabric defect detection	DeepLabV3+ and Efficient Net	98,777	public	fabric (swatch)	created
[44]	2024	Fabric defect detection	DL and SVM	3375	public	multiple	created
[45]	2020	Fabric defect detection	LSTM-CNN	12,000	public	fabric (swatch)	existing
[46]	2023	Fabric defect detection	LSTM-CNN	4100	private	fabric (swatch)	created
[47]	2019	Fabric defect detection	YOLO and CNN	245	public	fabric (swatch)	created
[48]	2024	Fabric defect detection	NN and SVM	1300	public	fabric (swatch)	created

and

Table 5. Overview of studies utilizing AI techniques for fabric surface assessment.

Paper	Year	Study Focus	Technique	No: Images	Dataset Type	Dataset Source	Fabric/Fiber Type
[49]	2021	Pilling assessment	CNN	25,876	fabric (swatch)	created	Woolen knitted
[50]	2019	Pilling classification	PCA and NN	1920	fabric (swatch)	existing	Knitted
[51]	2023	Pilling assessment	LSNet	42,304	fabric (swatch)	created	Woven, knitted, and nonwoven
[52]	2022	Pilling assessment	SONet	64,109	fabric (swatch)	created	Knitted, nonwoven fabric
[53]	2022	Pilling classification	Deep learning Net	320	fabric (swatch)	created	Knitted
[54]	2023	Pilling area segmentation	U-Net network	1000	fabric (swatch)	created	Not given
[55]	2020	Pilling detection	Cellular NN	400	fabric (swatch)	unclear	Cotton, hemp, and wool
[56]	2023	Pilling assessment	Saliency-based CNN	1650	fabric (swatch)	mixed	Woven fabrics
[57]	2022	Fiber flaw detection	CNN	5268	other	created	Not given
[58]	2024	Pilling assessment	UNet network	153	fabric (swatch)	created	Varying fiber contents, tissue structures, and colors
[59]	2020	Fabric smoothness appearance assessment	CNN	4620	fabric (section)	created	Not given
[60]	2021	Fabric surface assessment	Deep learning	7000	fabric (microscopic)	created	Not given

.

Table 3 focuses on studies related to textile sorting and material classification, highlighting AI techniques used to distinguish between fiber types, material blends, and garment categories. Table 4 presents research on garment and fabric defect detection, showcasing AI-driven approaches for identifying faults such as broken stitches, damage in zippers, or general wear-and-tear. Finally, Table 5 compiles studies addressing surface-level fabric assessments, including pilling detection, fiber flaw identification, and fabric smoothness evaluation. Each table includes detailed attributes such as the study’s focus, AI technique, dataset type, and fabric/fiber information. This structured layout facilitates clearer cross-study comparison and reveals domain-specific trends in dataset usage, model complexity, and fabric variability. Through this classification, we aim to enhance clarity and provide targeted insights into how AI is being leveraged across key aspects of clothing quality evaluation and textile reuse. For instance, across all three domains, CNNs are the most frequently used technique, consistently appearing in textile material recognition studies [19,23,26,27], fabric defect detection [30,36,38], and surface-level assessments such as pilling detection [49,56,57]. This widespread use is attributed to CNNs’ strong performance in image-based tasks including classification, detection, and segmentation. In addition to CNNs, YOLO (You Only Look Once) models such as YOLOv5 and YOLOv8 are commonly applied for fast and accurate object detection, particularly in garment defect detection scenarios [29,30,32,41]. Furthermore, several studies have employed hybrid models, such as CNN-LSTM for combining spatial and temporal features [22,45,46] and DeepLabV3+ integrated with EfficientNet for multi-scale feature extraction and efficient segmentation [42,43]. These models demonstrate the field’s shift toward increasingly sophisticated AI architectures tailored to specific challenges in textile analysis.

These tables also include information about the types of images within the datasets, such as what type of textiles were captured, which ranged from full garments to larger sections of fabrics, small switches of materials, and close-up or microscopic images of yarns and fabrics. The source of the images was also noted, as some studies sourced existing images while others created the images in their datasets, with both approaches often increasing the dataset size through image manipulations such as flipping or cropping.

The reviewed textile sorting studies utilized both natural and synthetic fabrics, with cotton, polyester, wool, silk, and viscose being the most common, along with various blended materials. Cotton-rich textiles appeared frequently, likely due to their high recycling value. Similarly, the reviewed studies on pilling identification (Table 5) predominantly utilized knitted fabrics, reflecting their common use in apparel susceptible to surface wear. Some studies extended their scope to include woven and nonwoven materials, while a few incorporated specific fibers such as wool, cotton, and hemp or considered variations in fiber content, fabric structure, and coloration to simulate real-world conditions. Nonetheless, several studies failed to report detailed fabric specifications, which poses a limitation for assessing the generalizability of AI models. This inconsistency underscores the necessity for standardized and comprehensive dataset documentation to facilitate robust model evaluation and applicability across diverse textile categories.

4. AI Applications and Their Impact

AI applications in garment defect detection have become indispensable in modern textile industries due to their ability to automate traditional manual processes, significantly improving accuracy, efficiency, and scalability.

4.1. Garment Defect Detection

4.1.1. Fabric Pilling Detection

Wu et al. [49] addressed the challenge of evaluating fabric pilling, a defect often assessed manually, leading to subjective and inconsistent results. They proposed a double-branch convolutional neural network (D-DCFNet) designed to objectively evaluate the degree of pilling in woolen knitted fabrics, semi-worsted knitted fabrics, and nonwovens using a pilling box method. Their model utilized a fusion of high-level features and cross-level representations, outperforming traditional inspection methods. However, the computational complexity of the model may restrict its adoption in resource-limited settings. Additionally, the dataset was limited to specific fabric types, potentially reducing the model’s generalizability to other textiles.

Yang et al. [50] introduced a deep principal component analysis-based neural network (DPCANN) for fabric pilling classification. This study focused on automating pilling identification, eliminating inefficiencies associated with traditional inspections. The methodology involved feature extraction using deep PCA followed by classification using either neural networks or support vector machines (SVMs). While computationally efficient, the reliance on smaller datasets which did not include details of the types of fabrics assessed could limit the model’s robustness in real-world applications, particularly for fabrics outside the study’s scope.

4.1.2. Stitching Defect Detection

Kim et al. [41] focused on detecting broken stitches during garment production. They developed a CNN-based stitching defect detection algorithm effective in scenarios where fabric and stitch colors are similar, a common challenge in edge detection. The dataset was sourced using low-cost Pi cameras with an 8-megapixel Sony IMX219 sensor and supplemented with online image repositories. Despite its innovative approach, the reliance on low-resolution cameras and limited dataset diversity (using only seven stitched samples and using a straight single-needle lockstitch) may hinder scalability in industrial environments.

Fang et al. [39] developed a two-stage zipper tape defect detection framework based on fully convolutional networks. In the first stage, they used a multi-scale detection architecture to efficiently identify large context regions likely to contain small-scale defects by fusing shallow and high-level features. In the second stage, they performed fine-grained detection of the small-scale defects within those regions. Their method demonstrated high accuracy, efficiency, and robustness compared to existing approaches.

4.1.3. General Garment Defects

Wu et al. [52] presented an advanced defect detection framework using LSNet, a CNN-based architecture tailored for multi-feature fusion. This model addressed common defects such as fuzzing and pilling, which account for a significant proportion of textile quality complaints annually. The dataset in this study comprised over 10,000 images of woven, knitted, and nonwoven fabrics, categorized into six fabric types. While achieving high accuracy in defect detection, the model’s dependency on high-quality labeled data may limit its applicability across diverse fabric types.

4.2. Textile Sorting and Recycling

4.2.1. Recycling-Oriented Sorting

Riba et al. [24] explored the integration of near-infrared (NIR) spectroscopy with CNNs for classifying post-consumer textile waste. Their system categorized textiles into seven distinct classes, including cotton, wool, and polyester, achieving a classification accuracy exceeding 90% for fabrics that were composed of 100% of a single fiber. Despite its effectiveness, the system struggled with sorting blended textiles and fabrics which is a strong limitation given most of the consumer fashion and textile products are composed of blended fiber types.

Kukreja et al. [22] developed a hybrid CNN-LSTM model for textile classification, categorizing fabrics into four recycling pathways: mechanical, chemical, upcycling, and downcycling. The dataset included 10,000 labelled images from second-hand stores and fabric shops, as well as online. While reporting high accuracy, the model’s reliance on balanced datasets may not reflect real-world imbalances in textile waste streams, which could limit its scalability in large-scale industrial recycling systems. However, the process of categorizing fabrics as mechanical, chemical, upcycling, and downcycling was not described in detail, and it is unclear what the criterion was for each. It was also unclear how the classification of the fabrics for the four different resulting pathways was determined, suggesting a knowledge gap between those investigating AI solutions and the practicalities of current system requirements within the F&T sector.

4.2.2. Garment Categorization for Reuse

Tian et al. [13] proposed an AI-driven machine vision system utilizing attention mechanisms for classifying post-consumer garments. The dataset comprised an initial 19,800 images across 11 garment categories, followed by a further 7033 for testing, ensuring robustness in real-world applications for sorting garments by archetype. The system achieved superior performance compared to conventional sorting methods but faced limitations in handling garments with overlapping features or heavily degraded fabrics and was unable to sort according to fiber type.

5. Assessing Dataset Availability and Practicality

The success of AI applications hinges on robust datasets. Several studies (as follows in the next subsections) have compiled extensive datasets tailored to specific applications in garment quality assessment and textile recycling.

5.1. Fabric Pilling Detection Datasets

Fabric pilling is a common issue in textiles that affects overall appearance and durability. Detecting and evaluating fabric pilling is crucial for quality control in the textile industry. The industry standards rely on human judges to rate the fabrics on a scale from one to five. The automation of pill detection by AI could facilitate a pilling rating system. Various datasets have been developed to aid in the detection and classification of fabric pilling, utilizing different methodologies and standards.

Some datasets are based on standard pilling images provided by institutions such as the Suzhou Institute of Science; however, they were created using multiple replicates of a single type of fabric, in this case a 20/80 polyester/cotton twill weave [56]. However, some datasets only examined a limited number of fabrics, e.g., in the case of the dataset in [58], in which four colored and two lightly colored knitted samples were used. No other fabric specifications such as fiber type were provided.

Another dataset [49] typically included images of woven fabrics with pilling levels ranging from 1 to 5, as per standards like GB/T 4802.1:2008 [61]. In addition to standard datasets, nonstandard pilling images are also collected using fabric pilling instruments like the YG502B. This study employed various data augmentation methodologies for generating “nonstandard” images, which are essential for training deep convolutional networks in pilling grade classification. Techniques such as multi-angle rotation (−45° to +45°), random cropping, random flipping (horizontal and vertical), and adding Gaussian noise are utilized. These methods simulate diverse fabric orientations, wear patterns, and real-world imperfections. Together, they significantly expanded the dataset to 1650 images, enhancing its representativeness of real scenarios without tampering with the number of pills in a certain area of the images [56].

With advancements in deep learning, various datasets have been developed to train models for automatic pilling detection and classification. For example, a dataset [54] was compiled to evaluate the effectiveness of a dual-attention U-Net model for fabric pilling detection. This dataset includes 100 different fabric images with pilling grades of 1–4 only. Fabrics with a pilling rating of ’5’ were excluded because they lacked any pills which in turn were marked. The 100 images were then used to generate a dataset of 1000 using a data augmentation strategy to increase the number of samples by applying various geometric transformations. Similarly, Yang et al. [50] utilized two datasets procured from a manufacturer located in Taiwan, which included fabrics with pilling ratings ranging from 2 to 5, according to ISO 12945-2:2020 [62], using the Martindale wear tester. The initial dataset encompassed 320 images, with 80 records allocated to each rating, whereas the subsequent dataset consisted of 1600 images, with 400 records corresponding to each rating. No other details of the fabrics were provided.

Some datasets were very large but with limited applications, such as the dataset used by Wu et al. [52] which included 64,109 images of woven, knitted, and nonwoven fabrics tested using the pilling box method. Whilst the dataset contained a very large number of images, the total number of different fabrics examined was only six in total, comprising either polyester or a wool or polyester blend, with around 10,000 images of each fabric. Similarly, Wu et al. [51] created another dataset of 42,304 pilling images of four different knitted and nonwoven wool and polyester fabrics using the pilling box method. This dataset also classified hairballs into large, small, and tangled configurations for a detailed analytical approach. In a similar vein, Wang et al. [55] concentrated on hairball detection, and while they used images of cotton, hemp, and wool fabrics, no other details about the fabrics or number of images were provided.

5.2. Sorting Datasets

The sorting of textiles and garments is a critical component in the reuse and recycling process, with the introduction of AI technology aiming to enhance efficiency and speed. Various datasets have been developed to facilitate the sorting of textiles, leveraging advanced technologies such as machine vision and near-infrared spectroscopy. These datasets are crucial for training and evaluating sorting systems, ensuring they can handle the complexities of textile waste, including variations in garment or item type, fiber type, color, construction, and condition.

Schrøder et al. [15] used a sample set of 123 pieces and a validation set of 154 pieces, consisting of both uniformly and multicolored textiles pre-sorted into six color categories (black, grey, red, white, blue, and ‘nature’). These textiles were manually pre-sorted by a waste textile sorting company. A critical review of the sample details would note the manual pre-sorting process, which introduces human error and subjectivity. The study’s reliance on manual sorting highlights a gap in automated sorting technologies, crucial for scaling up recycling efforts. Additionally, while the inclusion of both uniformly and multicolored textiles broadens the scope, it presents a challenge for consistency in sorting due to the inherent complexity of multicolored textiles. The sample size is moderate, allowing for some level of statistical analysis, but it may not fully represent the vast diversity of textile waste.

Various datasets have also been developed to identify the fiber content of textiles and garments, aiding in sorting for the purpose of recycling. For instance, Islam et al. [27] used 7000 garment images developed as part of a previous study [63] for detecting the percentage of cotton in textile samples. Similarly, Sun et al. [16] collected near-infrared (NIR) spectra for six different textile fibers, including cotton, viscose, acrylic, polyamide, polyester, and cotton–viscose blends for textile fabric classification. Once again, no other textile characterization details were provided (e.g., structure, thickness, or weight), and during testing parameters, it was specified that “a few” specimens were layered to prevent light leakage. In another study, 840 images of fabrics from the Contrado and The Textile District were collected at 50–100× magnification via a microscope attached to a smartphone [21]. The fabrics included categories according to tradename, such as Lycra, as well as fabric structure, i.e., crepe, jersey, and velvet, as well as fiber type, i.e., cotton–linen blend. These images are publicly available on Kaggle. Furthermore, Du et al. [26] established an online NIR spectral library, including 13 types of waste textiles identified only by fiber type, e.g., polyester, cotton, wool, and various blends such as silk/cotton and nylon/’spandex’. This library supports the development of NIR-based sorting systems by providing spectral data for different textile compositions. Similarly, another study [24] examined 370 commercially supplied textile samples, using NIR spectroscopy combined with deep learning to classify pure fibers and binary mixtures based on the fiber content information provided by the manufacturers.

5.3. Fabric Defect Detection

Fabric defect detection has emerged as a critical focus within textile manufacturing, responding to the increasing necessity for automated quality assurance in a sector that has predominantly depended on conventional manual visual inspection. A variety of datasets have been established to expand the research and practical applications in this area, each responding to specific challenges.

The ZJU-Leaper [43] initiative represents a comprehensive benchmark dataset comprising 98,777 images classified into 19 distinct fabric categories. This dataset introduces five distinct task configurations and an innovative evaluation protocol designed to reconcile the disparity between conventional inspection techniques and automated visual systems, highlighting the imperative for extensive, annotated datasets to propel advancements in defect detection technologies. As such, the paper addresses the need for efficient quality control in textile manufacturing using computer vision technology. It presents the ZJU-Leaper dataset and multiple-stage models for defect detection, emphasizing fast deployment and easy upgradability. The paper also introduces a new evaluation protocol to ensure accurate comparisons between different inspection methods. Similarly, ISL-Knit [44] specifically addresses knitted fabrics, bridging a significant gap by providing a dataset that includes 3018 defective and 357 flawless images. Through practical applications on devices such as Raspberry Pi, it demonstrates the practicality of deep learning in industrial quality control. Another dataset, Textil-5k [46], encompasses a compilation of 4100 images, where labelled and unlabeled data promote advancements in both supervised and semi-supervised learning, moving the industry closer to scalable, automated solutions.

Publicly annotated databases also aspire to standardize evaluation practices. For instance, Silvestre-Blanes et al. [47] presents a plain fabric dataset to address the inconsistencies associated with the utilization of private collections in verifying defect detection methodologies. By categorizing detection strategies and underscoring diverse imagery, it establishes a benchmark for future research evaluations. In a parallel context, a newly developed patterned texture [48] dataset tackles challenges related to anomaly detection, with methodologies such as reverse distillation demonstrating superior performance compared to traditional approaches in assessing a variety of defect types.

5.4. Stitching Defect Detection

Stitching defects are another area of focus within textile manufacturing, and similar to fabric or textile defects, the development of machine vision in this area is a response to the increasing necessity for automated quality assurance that has predominantly relied on manual visual inspection methodologies.

Few datasets containing images of stitch defects have been acknowledged by [42]. A dataset of 900 images of stitching defects was created to be analyzed using DeepLabV3+. Half of the samples were created with white thread on white fabric, with the rest created using blue thread on white fabric. No other details, such as the fabric types or the ISO stitch types or the machines used to create the stitches were provided. Focusing on image segmentation techniques, the paper employs a deep learning algorithm for defective stitch inspection in garment manufacturing. Specifically, their proposed DeepLabV3+ architecture, combined with EfficientNet models, is utilized to precisely identify and classify defective stitches.

6. Societal and Economic Impacts

AI-driven machine vision systems for textile assessment and recycling offer far-reaching societal and economic benefits. These technologies have the potential to optimize industrial processes, enhance the efficiency of second-hand markets, and support NGO operations. The following sections discuss the impact of adopting these systems on second-hand markets, charitable initiatives, and the efficacy of operations with cost benefits and the transition to more circular practices.

6.1. Second-Hand Markets

AI-based quality assurance systems have the potential to enable second-hand clothing collectors, sorters, and retailers to perform consistent and accurate inspections. By automating defect detection and quality grading, these systems could improve pricing accuracy, fostering consumer trust and streamlining the resale process, ensuring that only high-quality second-hand garments reach consumers.

6.2. NGO and Charitable Initiatives

For NGOs distributing clothing to underserved communities, AI-driven sorting systems ensure that donated garments meet usability standards. Riba et al. [17] highlighted the role of automated textile classification in reducing waste and enhancing transparency in donation systems. By efficiently separating reusable items from recycling streams, these systems empower NGOs to focus on their core missions.

6.3. Increased Efficiency

AI technologies can reduce operational costs by minimizing reliance on manual inspections and increasing sorting efficiency. Kukreja et al. [22] emphasized how hybrid models could optimize recycling pathways, enabling industries to recover maximum value from waste textiles while minimizing environmental impact. Such advancements align closely with circular economy goals, driving sustainability in the textile sector.

6.4. Enhancing Circular Economy Practices

The adoption of AI technologies is pivotal to realizing the principles of a circular economy in the textile and fashion industries. By minimizing waste, maximizing resource recovery, maximizing efficiency, and enabling the reuse of garments, AI offers a potentially transformative approach to the fashion and textiles industry, particularly the second-hand sector. Tian et al. [13] demonstrated how improvements in automated garment classification systems could enhance sorting accuracy for reuse, directly contributing to waste reduction and material conservation. This not only enhances resource efficiency but also aligns with global efforts to combat the environmental challenges posed by textile waste.

7. Challenges and Limitations

Despite the numerous benefits, AI technologies in garment quality assessment and textile recycling face several challenges:

7.1. Dataset Availability and Quality

Wu et al. [49] and Yang et al. [50] noted limitations in their dataset’s generalizability, as these were tailored to specific fabric types and conditions. Expanding dataset diversity to include a broader range of materials and garment states is essential for the scalability of AI solutions. For example, Tian, et al. [13] were able to work with a very large dataset; however, they were limited by the types of garments within the dataset and acknowledged the influence of a socio-economic bias on the garment dataset. Tian, et al. [13] reported on the difficulty of unbalanced datasets, requiring an equal number of images of each type of garment type to be identified to sufficiently train their model, and Sanjana [41] reported on the lack of datasets containing stitches and seams.

7.1.1. Textile Characterization

Many of the studies examined in this review did not provide sufficient details on the types of textiles included in their datasets. The variability amongst textiles in commercial use is almost infinite, so the details of the types of textiles included are essential information to understand the usefulness and applicability of the AI models. At the very minimum, the fabric structure and fiber types should be included, though ideally the characterization details of all fabrics within the dataset would be available (fiber, yarn, structure, mass per unit area, and thickness).

7.1.2. Textile Identification

Difficulties in recognizing certain visual aspects of textiles are persistent. While the color sorting system was 90% accurate, certain colors were still difficult for the system to recognize, and reflective textiles posed specific issues [15].

7.2. Computational Demand

Kukreja et al. [22] highlighted that their hybrid CNN-LSTM model required substantial processing power, potentially limiting its application in resource-constrained environments such as small-scale recycling facilities.

7.3. Environmental Variability

Studies by Riba et al. [24] and Luo et al. [36] pointed out challenges related to handling varying lighting conditions and image quality in real-world settings, which can hinder model performance and adoption.

7.4. Model Complexity

Advanced models like LSNet [51] and YOLO-SCD [36] exhibit high computational complexity, making them less accessible for industries with limited technological infrastructure.

7.5. Generalizability

Many models are trained on specific datasets and in the case of the textile datasets used only a limited number of textiles or did not provide sufficient detail on the types of textiles used, thus reducing their applicability across diverse textile types and real-world scenarios. Ensuring models can generalize across a variety of textile compositions (knits, wovens, and nonwovens of varying fiber types) and defect types remains a critical challenge.

7.6. Image Variety

There is a range of types of textile types studied (see Figure 5) using different types of imaging, ranging from microscopic images of textiles to hyperspectral images of yarns and textiles and whole garment photographs. These different image types highlight that existing datasets may provide only limited information for future studies, as defects can be observed as occurring across a range of aspects of the textiles, from fiber and yarn faults to fabric performance, localized areas of a garment or trim, as well as whole garment assessments such as categorization of garment types. The types of textiles captured were most commonly swatches of fabric, with 20 studies examining these, whereas whole garments are less frequently studied, with only 7 instances. Considering the increasing interest in AI for supporting textile sorting and the reuse of existing garments, studies which examine whole garments or parts of garments, rather than fabrics or yarns, provide a potential future priority area.

8. Future Direction

Building on the insights from current studies, future research into AI applications for the F&T industry should focus on several key areas to drive innovation and sustainability. Firstly, expanding dataset diversity is crucial. Developing comprehensive datasets that encompass a wider range of fabric types, defect categories, and real-world conditions will enhance the generalizability of AI models. This approach ensures that models are robust and applicable across various scenarios, ultimately improving their reliability and effectiveness [64]. Clothing and textiles have almost infinite variability in the types of fibers, yarns, fabric structures, and construction used; therefore, datasets need to include a wider range of clothing and textiles samples. In addition, generative AI models, such as Generative Adversarial Networks (GANs) and diffusion models, can be leveraged to synthesize realistic images of garments under various conditions. These generated images can simulate different types of wear, damage, or aging, providing valuable training data for AI systems tasked with assessing garment quality or predicting durability. This approach not only augments existing datasets but also helps in developing more accurate and resilient models for evaluating the condition of textiles in recycling and reuse contexts. Secondly, enhancing model efficiency is essential for broader adoption, especially in resource-constrained environments. Innovating lightweight and computationally efficient AI models can facilitate their use in diverse settings, from small-scale workshops to large industrial operations. This focus on efficiency not only reduces computational costs but also makes advanced AI technologies more accessible [65].

Integrating multimodal data represents another promising direction. By combining visual data with other modalities, such as spectral or tactile information, researchers can significantly improve defect detection and textile classification accuracy. This multimodal approach leverages the strengths of different data types to provide a more comprehensive analysis of textile properties [66]. Developing real-time processing capabilities is also a priority. Real-time AI systems capable of handling high-throughput textile processing without compromising accuracy can revolutionize the industry. These systems enable immediate detection and correction of defects, enhancing productivity and reducing waste [64]. Establishing collaborative platforms for dataset sharing and model development is vital for fostering innovation and standardization. Such platforms can bring together researchers, industry practitioners, and NGOs, facilitating the exchange of knowledge and resources. This collaboration can accelerate advancements and ensure that best practices are widely adopted [67] and help overcome identified issues such as dataset diversity.

Given the global scope of the problem of textile waste, a cooperative, global solution such as open-source datasets and AI models to more efficiently re-circulate textiles and garments is needed. Integrating sustainability metrics into AI models is imperative. By incorporating metrics that assess and optimize the environmental impact of textile recycling and reuse processes, AI can play a pivotal role in promoting sustainable practices. This integration helps in making informed decisions that balance efficiency with environmental responsibility [65]. Finally, whilst there is a great deal of potential for the application of machine vision and AI technology within fashion and textiles for a circular economy, there is a need to incorporate stronger fashion and textiles disciplinary expertise and perspectives to develop technology which supports not only sorting but also ensuring that the sorted materials align to the needs of industry, who will then use, reuse, remake or recycle these textiles. The incorporation of additional relevant characterization details such as textile fiber type and structure, garment type, construction techniques, etc., will ensure the usefulness and relevance of technology solutions for a circular economy. Undertaking a multi-disciplinary approach is suggested to improve the relevance of future studies.

9. Conclusions

Machine vision and AI technologies hold immense potential to transform the second-hand textile and clothing industries by enhancing garment quality assessment, streamlining textile sorting, and promoting sustainable recycling practices within a circular economy framework. This systematic review underscores the significant advancements made through CNNs, hybrid models, and machine vision systems in addressing key sustainability challenges.

Computer vision presents a transformative opportunity for revolutionizing the sorting and classification of textiles and garments, playing a pivotal role in the transition towards a more circular fashion industry. By automating and significantly enhancing the accuracy of sorting processes, computer vision can effectively address the inherent limitations of manual methods, leading to substantially higher recycling rates and the recovery of better-quality recycled materials. While the potential of this technology is immense, continued research and development efforts are essential to overcome the existing challenges associated with the complexity of textile waste and the need for robust and adaptable AI systems. Realizing the full transformative potential of computer vision in creating a sustainable and circular fashion industry necessitates close collaboration among all stakeholders, including researchers, fashion brands, textile manufacturers, technology providers, end-of-life providers, and policymakers.

The inherent complexity and variability of textile waste presents a significant hurdle to achieve circularity. This waste stream includes a wide array of fiber blends, colors, textures, and attached components, making it challenging for computer vision systems to accurately analyze and categorize everything. The sorting complexity arising from mixed-fiber textiles is a well-documented issue. Training robust AI models that can handle this variability requires large and diverse datasets encompassing the full spectrum of textile waste. Expanding the size and diversity of these datasets is crucial for improving the accuracy and reliability of AI models. Additionally, computer vision systems must be able to effectively handle variations in lighting conditions, fabric textures, and the diverse types of damage that can be present in textile waste. This SLR points out to the challenges related to dataset diversity, textile characterization, computational demands, and model generalizability persist, necessitating additional ongoing research and innovation, with a greater need for interdisciplinary work and the incorporation of fashion and textiles perspectives. By addressing these challenges, AI technologies can play a pivotal role in fostering a sustainable and equitable textile economy, reducing waste, and conserving valuable resources.